Skip to contentSkip to conversation

cxl/pci: Skip to handle RAS errors if CXL.mem device is detached

PATCH New Li Ming 1 participant 1 message Jan 29, 2024 CXL

lore
Share & tools
lore

Li Ming

author
#0 lore raw

PATCH v2 1/1 cxl/pci: Skip to handle RAS errors if CXL.mem device is detached

The PCI AER model is an awkward fit for CXL error handling. While the expectation is that a PCI device can escalate to link reset to recover from an AER event, the same reset on CXL amounts to a suprise memory hotplug of massive amounts of memory.

At present, the CXL error handler attempts some optimisitic error handling to unbind the device from the cxl_mem driver after reaping some RAS register values. This results in a "hopeful" attempt to unplug the memory, but there is no guarantee that will succeed.

A subsequent AER notification after the memdev unbind event can no longer assume the registers are mapped. Check for memdev bind before reaping status register values to avoid crashes of the form:

 BUG: unable to handle page fault for address: ffa00000195e9100
 #PF: supervisor read access in kernel mode
 #PF: error_code(0x0000) - not-present page
 [...]
 RIP: 0010:__cxl_handle_ras+0x30/0x110 [cxl_core]
 [...]
 Call Trace:
  <TASK>
  ? __die+0x24/0x70
  ? page_fault_oops+0x82/0x160
  ? kernelmode_fixup_or_oops+0x84/0x110
  ? exc_page_fault+0x113/0x170
  ? asm_exc_page_fault+0x26/0x30
  ? __pfx_dpc_reset_link+0x10/0x10
  ? __cxl_handle_ras+0x30/0x110 [cxl_core]
  ? find_cxl_port+0x59/0x80 [cxl_core]
  cxl_handle_rp_ras+0xbc/0xd0 [cxl_core]
  cxl_error_detected+0x6c/0xf0 [cxl_core]
  report_error_detected+0xc7/0x1c0
  pci_walk_bus+0x73/0x90
  pcie_do_recovery+0x23f/0x330

Longer term, the unbind and PCI_ERS_RESULT_DISCONNECT behavior might need to be replaced with a new PCI_ERS_RESULT_PANIC.

Fixes 6ac07883dbb5 ("cxl/pci: Add RCH downstream port error logging")

---
Changes in v2:
- Reword changelog for more context.(Dan)

---
 drivers/cxl/core/pci.c | 43 ++++++++++++++++++++++++++++++------------
 1 file changed, 31 insertions(+), 12 deletions(-)
drivers/cxl/core/pci.c +31 -12
index 6c9c8d92f8f7..480489f5644e 100644
@@ -932,11 +932,21 @@ static void cxl_handle_rdport_errors(struct cxl_dev_state *cxlds) { }
 void cxl_cor_error_detected(struct pci_dev *pdev)
 {
 	struct cxl_dev_state *cxlds = pci_get_drvdata(pdev);
+	struct device *dev = &cxlds->cxlmd->dev;
+
+	scoped_guard(device, dev) {
+		if (!dev->driver) {
+			dev_warn(&pdev->dev,
+				 "%s: memdev disabled, abort error handling\n",
+				 dev_name(dev));
+			return;
+		}
 
-	if (cxlds->rcd)
-		cxl_handle_rdport_errors(cxlds);
+		if (cxlds->rcd)
+			cxl_handle_rdport_errors(cxlds);
 
-	cxl_handle_endpoint_cor_ras(cxlds);
+		cxl_handle_endpoint_cor_ras(cxlds);
+	}
 }
 EXPORT_SYMBOL_NS_GPL(cxl_cor_error_detected, CXL);
 
@@ -948,16 +958,25 @@ pci_ers_result_t cxl_error_detected(struct pci_dev *pdev,
 	struct device *dev = &cxlmd->dev;
 	bool ue;
 
-	if (cxlds->rcd)
-		cxl_handle_rdport_errors(cxlds);
+	scoped_guard(device, dev) {
+		if (!dev->driver) {
+			dev_warn(&pdev->dev,
+				 "%s: memdev disabled, abort error handling\n",
+				 dev_name(dev));
+			return PCI_ERS_RESULT_DISCONNECT;
+		}
+
+		if (cxlds->rcd)
+			cxl_handle_rdport_errors(cxlds);
+		/*
+		 * A frozen channel indicates an impending reset which is fatal to
+		 * CXL.mem operation, and will likely crash the system. On the off
+		 * chance the situation is recoverable dump the status of the RAS
+		 * capability registers and bounce the active state of the memdev.
+		 */
+		ue = cxl_handle_endpoint_ras(cxlds);
+	}
 
-	/*
-	 * A frozen channel indicates an impending reset which is fatal to
-	 * CXL.mem operation, and will likely crash the system. On the off
-	 * chance the situation is recoverable dump the status of the RAS
-	 * capability registers and bounce the active state of the memdev.
-	 */
-	ue = cxl_handle_endpoint_ras(cxlds);
 
 	switch (state) {
 	case pci_channel_io_normal:
Signature · 1 line
2.40.1